Figure 4
From: Dense cellular segmentation for EM using 2D–3D neural network ensembles
CDeep3M segmentation comparison. Comparison between the CDeep3M segmentation tool and our lab’s (LCIMB) best segmentation algorithm for a binary cell/non-cell segmentation problem on our evaluation dataset. (a) Orthoslice of the ground truth binary segmentation of the evaluation dataset. (b) Segmentation using our lab’s (LCIMB) best 3D ensemble. (c) Probability map produced by the CDeep3M ensemble after training on our data for 30000 iterations. The probability map is a per-voxel probability that the voxel belongs to a cell region, and it must be thresholded to produce a segmentation. (d) Segmentation from the CDeep3M ensemble with the best tested threshold of 0.5. This resulted in an MIoU of 0.935, compared to 0.946 for the LCIMB segmentation. In addition to a slight improvement in MIoU statistic, the LCIMB segmentation does a much better job of preserving boundaries between adjacent cells.
